Accelerating vision foundation model for efficient medical image segmentation.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Segment Anything Model (SAM) shows promise for AI-based medical image segmentation.
• High computational costs of SAM hinder real-time applications due to Vision Transformer complexities.

Purpose of the Study:

• To accelerate SAM for medical image segmentation.
• To improve segmentation quality and reduce memory usage.

Main Methods:

• Proposed a CNN-assisted tuning strategy to enable SAM to process smaller inputs, reducing patches and memory.
• Introduced a token pausing strategy to skip computation for less informative patches, addressing redundancy.
• Combined both strategies for efficient, adaptable medical image segmentation.

Main Results:

• Achieved 12x faster processing compared to existing SAM-based methods.
• Demonstrated superior segmentation performance on Synapse and ACDC benchmarks.
• Reduced memory consumption significantly by processing smaller inputs.

Conclusions:

• Identified input resizing and uniform patch processing as limitations for SAM in medical imaging.
• Developed an efficient strategy integrating adapter-based tuning and token pausing.
• Enhanced throughput and preserved segmentation performance for medical applications.